Re: [PATCH] sched_groups are expected to be circular linked list, make it so right after allocation

[Igor Mammedov <imammedo@redhat.com>]

On 05/09/2012 01:52 PM, Peter Zijlstra wrote:
> On Wed, 2012-05-09 at 13:44 +0200, Igor Mammedov wrote:
>> This patch fixes only build_sched_groups path, but there is another fail path
>> that results in below OOPS.
>> build_overlap_sched_groups() may exit without setting groups and later it will crash
>> init_sched_groups_power as well.
>
> if that allocation fails? Or is there another fail path?

build_overlap_sched_groups(struct sched_domain *sd, int cpu)
...
                 if (cpumask_test_cpu(cpu, sg_span))
                         groups = sg;
...

above test fails and leaves local var groups set to NULL
and before exit there is:

   sd->groups = groups;

which resets sd->groups to NULL and I'm not sure if it is correct at all to skip this
assignment if groups == NULL.

>
>>   But I just don't know how to fix it, so I've just
>> posted partial fix that reduces crash frequency.
>
>
>> And I have to admit that
>> cpu_active_mask and siblings map are busted but we either should not exit from builder
>> funcs with NULL group or BUG there if it is impossible to come-up with sane group
>> for insane domain span.
>
> I'm perfectly OK with taking the machine down, provided we can output
> useful messages as to what is broken first..

-- 
-----
  Igor